Topic Battle

Artificial Intelligence systems can theoretically store unlimited information, constrained only by available server capacity and the electricity bills of data centres. GPT-4 was trained on approximately 570 gigabytes of text data, a figure that sounds impressive until one realises an elephant's brain weighs roughly 5 kilograms and contains rather more useful information about actual survival.

The AI's memory, however, suffers from a critical limitation: it cannot form new memories during conversation. Each interaction begins from the same fixed knowledge state, like meeting someone with anterograde amnesia at every dinner party. The elephant, by contrast, will remember your face, your smell, and whether you were kind — and adjust its behaviour accordingly, forever.

AI systems excel at narrowly defined problems with clear objectives and abundant training data. They can defeat world champions at chess and Go, generate plausible legal documents, and identify objects in photographs with superhuman accuracy — provided those objects resemble their training examples.

However, AI systems struggle profoundly with common sense reasoning, novel situations, and problems requiring genuine understanding rather than pattern matching. They cannot reliably determine whether a photograph depicts a real scene or is obviously impossible. They lack the contextual awareness an elephant demonstrates when deciding whether a new water source is safe. Speed of processing cannot compensate for absence of comprehension.

AI systems now influence billions of human decisions daily, from search results to loan approvals to medical diagnoses. They moderate social media content, translate languages in real time, and increasingly shape what information reaches human attention. This influence extends to every internet-connected device on the planet.

The speed and scale of AI's global reach is unprecedented. Within two decades, systems that began as academic curiosities now mediate human experience to a degree no previous technology has achieved. Whether this influence is beneficial remains vigorously debated, but its magnitude is undeniable.

Training a large language model requires approximately 1.3 million kilowatt-hours of electricity, equivalent to the annual consumption of 120 average British households. This investment produces a system that can write poetry, generate code, and occasionally claim to have emotions it cannot possibly possess.

The AI's learning methodology, known as supervised and reinforcement learning, requires millions of examples to achieve competence that an elephant acquires from watching its mother for an afternoon. Each capability improvement demands exponentially more computational resources, following a scaling curve that would concern any energy economist. The AI learns impressively, but at costs the biosphere cannot indefinitely sustain.

AI systems can generate text expressing emotions with impressive fluency. They can write sympathetic responses to human distress, compose love poetry, and simulate the language of grief. What they cannot do is actually feel anything. There is no experience behind the expressions, no qualia beneath the syntax.

This limitation becomes apparent in extended interaction. The AI's emotional responses are statistical averages of training data, not genuine reactions to present circumstances. It performs empathy without possessing it, demonstrates concern without experiencing it. The simulation of emotion, however convincing, differs categorically from emotion itself.